文档摘要:针对现有基于伪点云的3D目标检测算法精度远低于基于真实激光雷达(LightDetectionandranging,Li-Dar)点云的3D目标检测,本文研究伪点云重构,并提出适合伪点云的3D目标检测网络.考虑到由图像深度转换得到的伪点云稠密且随深度增大逐渐稀疏,本文提出深度相关伪点云稀疏化方法,在减少后续计算量的同时保留中远距离更多的有效伪点云,实现伪点云重构.本文提出LiDar点云指导下特征分布趋同与语义关联的3D目标检测网络,在网络训练时引入LiDar点云分支来指导伪点云目标特征的生成,使生成的伪点云特征分布趋同于LiDar点云特征分布,从而降低数据源不一致造成的检测性能损失;针对RPN(RegionProposalNetwork)网络获取的3D候选框内的伪点云间语义关联不足的问题,设计注意力感知模块,在伪点云特征表示中通过注意力机制嵌入点间的语义关联关系,提升3D目标检测精度.在KITTI3D目标检测数据集上的实验结果表明:现有的3D目标检测网络采用重构后的伪点云,检测精度提升了2.61%;提出的特征分布趋同与语义关联的3D目标检测网络,将基于伪点云的3D目标检测精度再提升0.57%,相比其他优秀的3D目标检测方法在检测精度上也有提升.
Abstract:Inviewoftheaccuracyofexisting3DobjectdetectionalgorithmsbasedonPseudo-LiDarisfarlowerthanthatbasedonrealLiDAR(LightDetectionandranging),thispaperstudiesthereconstructionofPseudo-LiDarandproposesa3DobjectdetectionalgorithmsuitableforPseudo-LiDar.ConsideringthatthePseudo-LiDARobtainedbyimagedepthisdenseandgraduallysparsealongtheincreaseofdepth,adepthrelatedPseudo-LiDARsparsificationmethodisproposedtoreducethesubsequentcalculationamountwhileretainingmoreusefulPseudo-LiDARinthemiddleandlongdistance,soastorealizethereconstructionofPseudo-LiDAR.Furthermore,a3Dobjectdetectionalgorithmbasedonobjectfeaturedistri-butionconvergenceundertheguidanceofLiDarpointcloudandsemanticassociationisproposed.Duringnetworktrain-ing,alaserpointcloudbranchisintroducedtoguidethegenerationofPseudo-LiDARobjectfeatures,sothatthegeneratedPseudo-LiDarobjectfeaturedistributionconvergestothefeaturedistributionoflaserpointcloudobject,therebycorrectingthedetectionerrorcausedbythedifferencebetweenthetwodatasources.AimingattheinsufficientsemanticassociationbetweenPseudo-LiDarinthe3Dcandidatebounding-boxobtainedbyRPN(RegionProposalNetwork)network,anatten-tionperceptionmoduleisdesignedtoembedthesemanticassociationbetweenpointsthroughtheattentionmechanisminthefeaturerepresentationofPseudo-LiDar,soastoimprovetheaccuracyof3Dobjectdetection.TheexperimentalresultsonKITTI3Dobjectdetectiondatasetshowwhentheexisting3DobjectdetectionnetworkadoptsthereconstructedPseudo-LiDar,thedetectionaccuracyisimprovedby2.61%.Furthermore,theproposed3Dobjectdetectionnetworkwiththefea-turedistributionconvergenceandsemanticassociationimprovestheaccuracyby0.57%.Comparedwithotherexcellentmethods,italsoimprovesthedetectionaccuracy.
作者:郑锦 蒋博韬 彭微 王森 Author:ZHENGJin JIANGBo-tao PENGWei WANGSen
作者单位:北京航空航天大学计算机学院,北京100191;虚拟现实技术与系统全国重点实验室,北京100191北京航空航天大学计算机学院,北京100191
刊名:电子学报 ISTICEIPKU
Journal:ActaElectronicaSinica
年,卷(期):2024, 52(5)
分类号:TP391.4
关键词:3D目标检测 伪点云 语义关联 分布趋同 注意力感知
Keywords:3Dobjectdetection Pseudo-LiDar semanticassociation distributionconvergence attentionperception
机标分类号:TP391TN914.5TP273
在线出版日期:2024年7月22日
基金项目:LiDar点云指导下特征分布趋同与语义关联的3D目标检测[
期刊论文] 电子学报--2024, 52(5)郑锦 蒋博韬 彭微 王森针对现有基于伪点云的3D目标检测算法精度远低于基于真实激光雷达(LightDetectionandranging,Li-Dar)点云的3D目标检测,本文研究伪点云重构,并提出适合伪点云的3D目标检测网络.考虑到由图像深度转换得到的伪点云稠密...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
关键词:3D目标检测,伪点云,语义关联,分布趋同,注意力感知,
|
|